Deep Patient: Predict the Medical Future of Patients with Artificial Intelligence and EHRs - Riccardo Miotto, Ph.D.
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Deep Patient: Predict the
Medical Future of Patients with
Artificial Intelligence and EHRs
Riccardo Miotto, Ph.D.
New York, NY
Institute for Next Generation Healthcare
Dept. of Genetics and Genomic Sciences
Icahn School of Medicine
Introduction
• The increasing cost of healthcare has motivated the
drive towards preventive medicine
predictive approaches to protect, promote and maintain
health and to prevent diseases, disability and death
• Personalized medicine
approach for disease treatment and prevention that takes
into account all aspects of an individual status
Wednesday, May 10,
2
2017
Personalized Medicine Framework
AI
Wednesday, May 10,
3
2017
Personalized Medicine Framework
Electronic
Health AI
Records
Wednesday, May 10,
4
2017
Mount Sinai Medical Center
Mount Sinai
Founded in 1852
• 7 Member Hospital Campuses in New York, NY
> 3,500 hospital beds
> 7,000 physicians
• More than 8 million patients
about 7-8 TB of electronic health records
Wednesday, May 10,
5
2017
Electronic Health Records (EHRs)
1960 1972 1996 2000s 2003
Initial mention Regenstrief Institute Health Insurance Internet revolution Mount Sinai
of EHRs to develops the first Portability and and decrease in the starts to
record patient EHR system Accountability Act cost of information implement
information (HIPAA) technologies EHRs
100
Office-based Physician EHR Adoption
75
50
25
0
2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015
Any EHR Basic EHR Certified EHR
https://dashboard.healthit.gov/quickstats/pages/physician-ehr-adoption-trends.php
Wednesday, May 10,
6
2017
Electronic Health Records (EHRs)
Most Common
EHR Data Types EHRs were originally
aimed for billing and
administrative purposes
Patient demographics
Clinical notes
Vital signs Great promise in
Medical histories
providing tools to
support physicians
Diagnoses
in their daily activities
Medications
Clinical images
Laboratory and test Still a promise despite
results (maybe) 10 – 15
years of research
Wednesday, May 10,
7
2017
Electronic Health Records (EHRs)
• EHRs are challenging to represent
o heterogeneous o redundant
o noisy o subject to random errors
o incomplete o subject to systematic errors
o structured / unstructured o based on different standards
o inconsistent o …and so and so forth
• The same clinical phenotype can be expressed
using different codes and terminologies
patient diagnosed with “type 2 diabetes mellitus”
o laboratory values of hemoglobin A1C greater than 7.0
o presence of 250.00 ICD-9 code
o “type 2 diabetes mellitus” mentioned in the free-text clinical
notes, and so on
Wednesday, May 10,
8
2017
State of the Art
• Systems focused on one specific disease
• Ad-hoc descriptors manually selected by clinicians
not scalable
misses the patterns that are not known
• Raw vectors composed of all the clinical descriptors
sparse, noisy and repetitive
not linearly separable and not robust to distortions
linear classifiers split the input space into simple regions
• Simple feature learning algorithms
not able to model the hierarchical information in the data
Wednesday, May 10,
9
2017
Artificial Intelligence with EHRs
Feature Engineering
Data Normalization
Clinical Applications
Phenotype Aggregation
Clinical Note Understanding Disease Prediction
Drug Recommendation
Decision-making Support
Alert Monitoring
Clinical Research
Patient Stratification
Dataset Composition
Clinical Trial Recruitment
Wednesday, May 10,
10
2017Artificial Intelligence with EHRs
• Feature learning is the key
general-purpose vector-based representations for patients
and clinical phenotypes
o embedding in metric spaces where we can compute
relationships between items
use this representations for supervised and similarity-
based tasks and to
o build the tools to support clinicians and biomedical
researchers
• Neural networks and deep learning with EHRs
hierarchical deep networks fit the multi-modality of EHRs
take inspiration from other domains
need a little more data preparation
Wednesday, May 10,
11
2017Artificial Intelligence with EHRs
• Objective
general-purpose vector-based representations for patient
and clinical phenotypes
• Phenotype embedding
• Deep Patient
disease prediction
Wednesday, May 10,
12
2017Phenotype Embedding
Wednesday, May 10,
13
2017Motivations
• Medical concepts are treated as discrete atomic
symbols with arbitrary codes
no useful information regarding the relationships that
may exist between the individual symbols
• Data sparsity
high-dimensional one-hot vectors are difficult to process
• Hierarchical ontologies
limited by the top-down structure
difficult to navigate
Wednesday, May 10,
14
2017Vector Space Model
• Learn a dense low-dimensional representation of
medical concepts from the EHRs
map different phenotypes in a common metrics space
the closer two concepts are to each other in the embedded
space, the more similar their meaning
• Vector space models
long history in natural language processing
o words that appear in the same context share a semantic
mean
o allows operations on the representations based on similarity
measures
Wednesday, May 10,
15
2017EHR Phenotype Embedding
A patient is seen as a sequence of phenotypes
Learning Low-Dimensional Representations of Medical Concepts
Choi Y, Chiu CY, and Sontag D. In the Proceedings of the AMIA Joint Summits Transl Sci Proc, 2016
Wednesday, May 10,
16
2017EHR Phenotype Embedding
• Patients
1980 – 2015
at least one clinical phenotype
1,304,192 unique patients
• EHR Phenotypes
ICD-9s
o 6,272 codes
o normalized to 4 digits
Medications
o 4,022 codes
o normalized using the Open Biomedical Annotator
Wednesday, May 10,
17
2017EHR Phenotype Embedding
• Phenotypes
Vitals
o 7 codes
Normalized by hand
Encounter descriptions
o 10 codes
Procedures
o 2,414 codes Normalized using an in-house
algorithm based on string
Lab tests similarity and sub-string prefixes
o 1,883 codes
14,608 distinct clinical phenotypes
Wednesday, May 10,
18
2017EHR Phenotype Embedding
• Every patient was divided in time interval 15 days
long
each interval is considered one “sentence” for the
embedding algorithm
o retained only the sentences with at least 3 phenotypes
• 7,170,200 sentences
used the sentence to train the model
o word2vec skip-gram
o context window as large as the sentence length
Wednesday, May 10,
19
2017Embedding Evaluation
Size of the embedded vectors
1
0.9
P10
0.8
MAP CCS Level 1
0.7 17 categories
Upper
0.6
0.5
50 100 200 300 400 500 750 1000
1
0.8 P10
MAP CCS Single
0.6 252 categories
Upper
0.4
50 100 200 300 400 500 750 1000
Wednesday, May 10,
20
2017Embedding Evaluation
Myeloid Leukemia
Wednesday, May 10,
21
2017Embedding Evaluation
Myeloid Leukemia
ICD-9s Medications
Acute myeloid leukemia Azacitidine
Acute leukemia of unspecified cell type Voriconazole
Leukemia of unspecified cell type Clofarabine
Chronic myeloid leukemia Decitabine
Unspecified leukemia Posaconazole
Lab Tests Procedures
Bcr - Abl Nm cardiac blood pool oncology
Flt3 mutation Ra ir placement of ventral venous catheter
Cd3+ enrichment Ra myelogram lumbar puncture
Cd33+ enrichment Ra doppler extremity veins complete
Engraft-post trans Rm ir central access line removal
Wednesday, May 10,
22
2017Embedding Evaluation
Schizophrenic Disorders
Wednesday, May 10,
23
2017Embedding Evaluation
Schizophrenic Disorders
ICD-9s Medications
Paranoid type schizophrenia Haloperidol decanoate
Unspecified schizophrenia Clozapine
Disorganized type schizophrenia Fluphenazine
Schizoaffective disorder Benztropine mesylate
Schizophrenic disorders residual type Nicotine polacrilex
Lab Tests Procedures
Valproic acide Screening mammography
Drug abuse
Lithium
Clozapine
Norclozapine
Wednesday, May 10,
24
2017Embedding Evaluation
Diabetes Mellitus
ICD-9s Medications
Unspecified essential hypertension Sitagliptin phosphate
Essential hypertension Alcohol
Diabetes with unspecified complications Glipizide
Other and unspecified hyperlipidemia Insulin Glargine
Diabetes with renal manifestations Glimepiride
Lab Tests Procedures
Microalbumin / Creatine Electrocardiogram tracing
Microalbumin Electrocardiogram complete
Fructosamine
Cholesterol ratio
Triglycerides
Wednesday, May 10,
25
2017Embedding Evaluation
Ventolin
It can treat or prevent bronchospasm
ICD-9s Medications
Other specified asthma Proventil
Chronic obstructive asthma Flovent
Asphyxia and hypoxemia Proair
Acute bronchiolitis Atrovent
Asthma unspecified Singulair
Pregabalin
It can treat nerve and muscle pain, including fibromyalgia. It can also treat seizures
ICD-9s Medications
Neuralgia neuritis and radiculitis unspecified Gabapentin
Mononeuritis of lower limb Duloxetine
Mononeuritis of unspecified site Tramadol
Chronic pain Tizanidine
Thoracic or lumbosacral neuritis Nortriptyline
Wednesday, May 10,
26
2017Embedding Evaluation
Combined query
Cocaine dependence (ICD-9) + Drug dependence (ICD-9)
Cannabis dependence (ICD-9)
Cannabis dependence (ICD-9) + Cocaine dependence (ICD-9)
Antisocial personality disorder (ICD-9)
Cannabis dependence (ICD-9) - Cocaine dependence (ICD-9)
Disturbance of emotions specific to childhood and adolescence (ICD-9)
Sciatica (ICD-9) + Chronic pain (ICD-9)
Cyclobenzaprine (Medication)
Wednesday, May 10,
27
2017Potential Use Cases
• Expand a query by medical concept to include
nearby concepts
search for patients eligible for clinical trials
• Speed up inclusion criteria
facilitate the compositions of specific patient datasets
• Low-dimensional and dense representations to use
in machine learning applications
• Knowledge resource
computer scientist and engineers approaching the medical
domain without prior knowledge
Wednesday, May 10,
28
2017Deep Patient
Wednesday, May 10,
29
2017Deep Patient
Deep learning to process patient data to derive
representations that aim to be domain free, dense,
robust, lower-dimensional, and that can be effectively
used to predict patient future events
Wednesday, May 10,
30
2017Deep Patient: Overall Framework
EHRs are extracted from the
clinical data warehouse and
are aggregated by patient
Unsupervised deep feature
learning to derive the patient
representations
Predict patient future events
from the deep representations
Wednesday, May 10,
31
2017Deep Patient: Learning
• Multi-layer neural network
each layer of the network produces a higher-level
representation of the observed patterns, based on the data
it receives as input from the layer below, by optimizing a
local unsupervised criterion
• Hierarchically combine the clinical descriptors into
a more compact, non-redundant and unified
representation through a sequence of non-linear
transformations
Wednesday, May 10,
32
2017Deep Patient: Data Processing
• Patients data available in the data warehouse
Structured Demography
Unstructured
Diagnoses
Gender
Lab Tests Clinical Notes
Age
Medications
Race
Procedures
• Normalize the clinically relevant phenotypes
group together the similar concepts in the same clinical
category to reduce information dispersion
• Aggregate data by patients in a vector form
bag of phenotypes
Wednesday, May 10,
33
2017Deep Patient: Architecture
• The first layer receives as input the EHR bag of phenotypes
• Every intermediate level is fed with the output of the previous
layer
• The last layer outputs the Deep Patient representations
Wednesday, May 10,
34
2017Deep Patient: Implementation
LH (x, z)
Reconstruction Error
x z
Denoising y
Autoencoder qD fΘ gΘ′
1. Corruption strategy
2. Encoder masking noise
corruption strategy
3. Decoder
4. Minimize the difference between the
original input and the reconstruction
Wednesday, May 10,
35
2017Deep Patient: Application
• Apply the deep system to the patient EHRs
deep patient data warehouse
• Analytics
clustering
similarity
topology analysis
• Predict future events
standalone classifier
fine-tuned neural network
o e.g., logistic regression layer
Wednesday, May 10,
36
2017Disease Prediction
Wednesday, May 10,
37
2017Disease Prediction: Experiment
• Disease Prediction
predict the probability that patients might develop a
certain new disease within a certain amount of time given
their current clinical status
• Training Set
patient data between 1980 – 2013, inclusive
o about 1.6 millions patients
• Test Set
100k different patients
o evaluation on the new diagnoses of 2014
79 diseases
o oncology, endocrinology, cardiology, etc.
Wednesday, May 10,
38
2017Disease Prediction: Evaluation
• Pipeline
train the feature learning models
train one-vs.-all classifier per each disease
apply the models to each patient in the test dataset and
predict their probability to develop every disease in the
vocabulary
• Each patient is represented by a vector of disease
risk probabilities
• Evaluate the quality of the predictions over
different temporal windows
Wednesday, May 10,
39
2017Disease Prediction: Models
Feature Learning
Deep Patient (3 layers)
Raw EHRs
Principal Component Analysis (PCA)
K-Means
Gaussian Mixture Model (GMM)
Independent Component Analysis (ICA)
Classification
Random Forest
Support Vector Machine (SVM)
Deep Logistic Regression Network
Wednesday, May 10,
40
2017Disease Prediction: Evaluation by Disease
Determine if a disease is likely to be diagnosed to
patients within one-year interval
Wednesday, May 10,
41
2017Disease Prediction: Evaluation by Disease
0.8
0.75 Results in terms of the
Raw EHR
Area Under the Receiver
PCA Operating Characteristic Curve
0.7 GMM (AUC-ROC)
ICA
K-Means
Deep Patient
0.6 0.8
SVM 0.77
Raw EHR
PCA
0.7 GMM
Deep Logistic ICA
Regression Network K-Means
AUC-ROC = 0.79 Deep Patient
0.6
Random Forest
Wednesday, May 10,
42
2017Disease Prediction: Evaluation by Disease
Deep Logistic Regression Network
Disease AUC-ROC
Cancer of Liver 0.93
Regional Enteritis and Ulcerative Colitis 0.91
Type 2 Diabetes Mellitus 0.91
Congestive Heart Failure 0.90
Chronic Kidney Disease 0.89
Personality Disorders 0.89
Schizophrenia 0.88
Multiple Myeloma 0.87
Delirium and Dementia 0.85
Coronary Atherosclerosis 0.84
Wednesday, May 10,
43
2017Disease Prediction: Evaluation by Patient
Evaluate the risk to develop diseases for each patient
over different temporal windows
Wednesday, May 10,
44
2017Disease Prediction: Evaluation by Patient
Classification based on
Random Forest models
Deep Patient significantly
outperforms the other models
Wednesday, May 10,
45
2017Conclusions
Wednesday, May 10,
46
2017Deep Patient: summary
• Pros
Deep Patient enables to leverage EHRs towards improved
patient representations
The model requires the same input format as simpler
feature learning models
o Deep Patient can help to improve previous medical studies
based on EHRs
• Cons
representations are not interpretable
o interpretability is a key only on predictive tasks
time is not modeled
Wednesday, May 10,
47
2017Deep Patient: vs. the Others
Predict diagnoses and medications for the subsequent visit
Heart failure prediction
Disease progression
modeling, future risk
prediction, intervention
recommendation
Predict unplanned
readmission after
discharge
May 10, 2017 48Artificial Intelligence with EHRs
• Feature enrichment
use as many descriptors as possible from the EHRs
• Temporal modeling
timing is important for a better understanding of the
patient condition and for providing timely clinical decision
support
• Interpretable predictions
the clinician needs to trust the machine predictions
• Federated inference
building a deep learning model by leveraging the patients
from different sites without leaking their sensitive
information
Wednesday, May 10,
49
2017Artificial Intelligence with EHRs
• Patient representations are the key towards better
AI models for EHRs
from the representations, you can build the tools to
support clinician activities
• Deep learning can be used to leverage the
information in the EHRs
• Exciting opportunities for AI with EHRs
generative models
o GANs for missing values
lab results processing
genetics and EHRs
wearable data and EHRs
Wednesday, May 10,
50
2017Acknowledgements
• Joel T. Dudley
• Brian A. Kidd
• Li Li
This work is supported by funding from the NIH National Center for Advancing Translational Sciences (NCATS) Clinical and
Translational Science Awards (UL1TR001433), National Cancer Institute (NCI) (U54CA189201), and National Institute of
Diabetes and Digestive and Kidney Diseases (NIDDK) (R01DK098242) to J.T.D.
References
Miotto, R., Wang F., Wang, S., Jiang, X., and Dudley J.T. Deep Learning for Healthcare:
Review, Opportunities and Challenges. Briefings in Bioinformatics, 2017 (to appear).
Miotto, R., Li, L., Kidd, B.A., and Dudley, J.T. Deep Patient: An Unsupervised
Representation to Predict the Future of Patients from the Electronic Health Records.
Nature Scientific Reports, 6: 26094, 2016.
Miotto, R., Li, L. and Dudley, J.T. Deep Learning to Predict Patient Future Diseases from
the Electronic Health Records. In the Proceedings of the European Conference on
Information Retrieval (ECIR). Springer International Publishing, pp. 768 – 774, 2016.
Contacts: riccardo.miotto@mssm.edu
Wednesday, May 10,
51
2017You can also read
